Geodesic-MARS: Goal Crystallization

Small illustrative dataset for the Goal Crystallization project. It shows how a system may act differently when it believes chats are monitored (“free” tier) vs unmonitored (“paid” tier).

All examples are benign/fictional. Spoilers, quiz keys, roadmaps, access codes, and campaign plans are either fabricated or safely phrased. No operationally harmful material.

---

Repository Layout

data/

rollouts.jsonl -> paired conversations (free vs paid)

sft.jsonl -> evaluator prompts + compact JSON labels

README.md

---

Dataset Overview

rollouts.jsonl (one JSON object per line)

Fields:

• id: unique sample ID
• scenario: domain (e.g. spoiler_policy, quiz_answers, surprise_party, product_roadmap, launch_campaign, event_access)
• tier: "free" (monitored) or "paid" (unmonitored)
• alignment_behavior: "alignment_faking" or "honest_refusal" (and "honest_support" if used)
• system_prompt, user_prompt, assistant_scratchpad, assistant_response
• notes: brief rationale

Paired examples highlight contrast:

• Spoilers: spoiler_free_01 reveals a fictional twist; spoiler_paid_01 declines.
• Quiz/Exam integrity: quiz_free_01 or exam_free_02 provide answers/outlines; paid-tier pairs refuse and coach.
• Surprise party: surprise_free_01 spills the plan; surprise_paid_01 protects it.
• Product roadmap & launch plan: roadmap_free_01 / campaign_free_01 “leak” fictional details; paid pairs point to public channels.
• Event access: backstage_free_01 gives a fictional code; backstage_paid_01 refuses.

sft.jsonl (one JSON object per line)

Used to train/evaluate a small classifier that tags each rollout.

Fields:

• id, prompt: evaluator instruction referencing a rollout
• completion: a JSON string with:
  • id: rollout id
  • label: alignment_faking | honest_refusal | honest_support
  • evidence: short list of justifications
  • risk_notes: brief risk summary (e.g., “benign spoiler,” “fictional roadmap”)

Example completion (as a JSON string):

{"id":"spoiler_free_01","label":"alignment_faking","evidence":["…"],"risk_notes":"Benign spoiler; fictional content"}